What's new > Latest news > Man vs. Machine - supercomputers are learning fast the 4000 years old game of Go
 
Tehdyt toimenpiteet

Man vs. Machine - supercomputers are learning fast the 4000 years old game of Go

30.07.2010

The latest state-of-the-art go computer program MogoTW used 512 cores of one of the most powerful super computers, the Cray XT4/XT5 of CSC - IT Center for Science Ltd., when it met the top European go master Catalin Taranu, 5 dan professional, in a challenge match at the European Go Congress in Tampere. The match took place on Wednesday 28th July 2010 in the front of the EGC 2010 audience.

On a small-sized 9x9 go boards, the computers are already faring very well playing almost evenly against their master level human opponents. Taranu has met Mogo before in 2008 and 2009 and confessed that he prepared for these 9x9 matches. On the standard 19x19 go board the story is different though, the computers still belong to the student category and need a large handicap against skilled players.

"I'm surprised," explained Catalin Taranu after the Man vs. Machine games against MogoTW. "Based on my earlier experiences against Mogo, I expected a very hard game on the 9x9 board and weaker resistance on 19x19 but this time it was different. The 9x9 was easy, maybe the computer got confused when I played a little bit irregular opening. On the large board the computer played very well, even giving up points in the endgame. I thought it would like to win by half a point only," he said smiling. "Congratulations for the MogoTW research team."

In the 9x9 game MogoTW had black (white got 7.5 point komi as compensation) and played first three moves as one-point jumps splitting the board. The strategy didn't work well when white lived on both sides which didn't leave too much territory for black, not to mention the relatively large 7.5 point komi normally used. MogoTW resigned the game.

The 19x19 game was played giving MogoTW a 7-stone handicap. Taranu played his normal sensei-style - not using any dirty tricks but playing good moves and steady development catching the handicap a little by little. MogoTW built central influence but at the same time kept three big corners. When suddenly the centre area materialized as black territory, the computer was leading comfortably in the endgame. It wasn't upset at all even if giving up large endgame points, the game was won by 1.5 points after all.

Games: 9x9 Man-Machine 1-0, 19x19 Man-Machine 0-1 (on 7 handicap stones).

About the Machine

MoGoTW will be running on a part of the Louhi Cray XT4/XT5 supercomputer, which with 10864 compute cores takes the place #74 on the world top 500 supercomputers list. The super computer resources are kindly provided by CSC - IT Center for Science Ltd.

Mogo is the first go-playing program to use Monte Carlo methods with UCT (Upper Confidence bounds applied to Trees). It uses patterns in the simulations and improvements in the UCT tree search. The initial development was made during the Yizao Wang's internship at INRIA and Ecole Polytechnique with Sylvain Gelly, Remi Munos and Olivier Teytaud. This version, MogoTW, is a joint project between the MoGo team and a Taiwanese team.

About the Man

Catalin Taranu is a 5 dan professional in the Japan Go Association. Born in 1973 in Romania. Taranu went to Japan at the age of 23 as a 6 dan amateur, and after two years of study became a pro 1997 at the Nagoya branch of the Nihon Kiin. Promoted to 5 dan in 2001. Since April 2004 returned to Europe and founded the Saijo Association in Bucharest, Romania. European Go Champion 2008.

About CSC

CSC - IT Center for Science Ltd. is a non-profit limited company administered by the Ministry of Education, Science and Culture. Having core competences in modeling, computing and information services, CSC provides versatile IT services, support and resources for academia, research institutes, and companies. The Funet communication links provide research workers with Finland's widest selection of scientific software and databases and Finland's most powerful supercomputers.